Vihan Patil

An agentic RAG health coach that only says what it can cite. A Corrective RAG (CRAG) loop re-retrieves when the evidence check comes back weak — and refuses outright when a second pass still can't support an answer. The citation ships with the answer, and an Android client feeds physiological data into the coaching context.

The interesting problem wasn't generating nutrition advice — any model does that. It was the corrective loop: when the evidence check finds too little support, the system rewrites the query and re-retrieves before drafting — and when a second pass still comes back weak it refuses rather than shipping an unsupported answer; the citation stays visible so a user can check the claim themselves. Responses are grounded in a curated medical-research corpus and evaluated for relevance, groundedness, and verifiability. Built as my M.S. capstone, advised by Prof. Leilani Gilpin of the AIEA Lab.

• Python
• RAG
• Verification Loops
• Health Data Sync
• LLM Orchestration

View repository ↗

An autonomous crop-survey drone in simulation. It breaks off mid-survey to dodge a bird nobody planned for, resumes the sweep, and reconciles the coverage ledger so nothing goes silently unscanned. NDVI crop-health mapping now runs from the same flight data.

The takeover is a real ArduPilot mode transition — AUTO to GUIDED and back — with a geofence-vetted 3D dodge, running live on ArduPilot SITL + Gazebo + ROS 2. The latest end-to-end flight met one bird mid-lane and corrected across four divert ticks in 0.43 s of GUIDED authority; 34 cells went at-risk, every one was re-imaged by a later lane, and the ledger still closed at 720 of 720. Getting there meant diagnosing a Fast DDS shared-memory bottleneck that was capping image recording at 0.41 Hz; it records at 5 Hz now. An independent ground-truth check then exposed a 6.7 mm bird clearance on a run that otherwise looked perfect, so that run doesn't count as a safety result and the main branch stays red until the fix is real. Replaying the encounter across 81 mission speeds and three aircraft models showed the nadir camera can never see far enough ahead — the case for forward-looking sensing, made with numbers. The demo below is the same mechanic in plan view — drag the bird and watch the executor pick a vetted dodge, hand authority back to AUTO, and reconcile the ledger.

• ROS 2
• Gazebo
• ArduPilot SITL
• Python
• Path Planning
• OpenCV

View repository ↗

An NLP research pipeline over SEC EDGAR: 1,271 filings from 25 mega-cap companies distilled into a labeled corpus for financial text-signal research, then walk-forward backtested against a numeric-only baseline — with the metadata discipline to keep every label honest about what the market knew, and when. The result was a null — and it's reported as one.

Financial text is booby-trapped. Boilerplate repeats filing after filing, and any label that peeks past the filing date poisons the dataset. The labeling pass — 6,746 chunks across sentiment, a red-flag taxonomy, guidance direction, and a distress tier — ran through the Anthropic Batch API behind deduplication and look-ahead-safe prompts, then a 400-chunk blind second-rater audit: 94–95% agreement on three of the four label families, 63% on red flags, kept as a limitation rather than ground truth. The punchline came from six chronological walk-forward folds of XGBoost, fundamentals-only against fundamentals-plus-text: the text gave no consistent lift once near-duplicate filings were controlled. The honest finding is the null result, and that's what the repo reports.

• Python
• NLP
• SEC EDGAR
• Anthropic Batch API
• Bootstrap Labeling
• XGBoost
• pandas

View repository ↗

Resume-to-job matching that explains itself: which requirements are covered, which are thin, and which are missing outright — with the exact phrases that earned each point of the score.

A single match percentage tells you nothing actionable — the useful output is which requirement you failed, and the evidence behind it. Scoring is deterministic, with no generative model in the path: documents are parsed on the recruiter's machine into SQLite, every score carries the exact phrases that earned it, and nothing leaves the device. The scoring core sits behind 1,100+ automated tests — property-based, metamorphic, adversarial — and an 80% Stryker mutation-score ratchet; a findings-driven release gate blocks any known wrong-score defect from shipping, because a confident wrong number about a real person is the one failure this project treats as unacceptable. The widget below runs the engine's own arithmetic client-side on synthetic fixture text — the cascade, the weights, the rounding, and the refusal.

• TypeScript
• Local-first
• SQLite
• pnpm monorepo

View repository ↗

Real-time on-chain visualization and smart-contract analysis, built with a team for UXly and shipped to production on EC2. It started as a UCSC course project; the client kept us on and scoped the requirements directly.

The React front end was the easy half. The real work was an ingestion layer that keeps up with block production, degrades gracefully when an upstream node stalls, and doesn't lose events while reconnecting. The engagement ran as weekly agile sprints with UXly's CEO and engineers — rotating scrum lead — and shipped to production on AWS EC2.

• TypeScript
• React
• Node.js
• Express
• Docker
• AWS EC2